Skip to content

Guest Lecture 1: Generative Reconstruction of Dynamic 3D Scenes in the Wild

Jiahui Huang

Senior Research Scientist at NVIDIA
April 21, 2025 (Mon), 1:00 p.m. KST
Online (Zoom).

Guest Lecture at CS479: Machine Learning for 3D Data
Minhyuk Sung, KAIST, Spring 2025


Teaser1

Abstract

Reconstructing 3D scenes is a fundamental task with broad applications in autonomous driving, robotics, and immersive visualization. While traditional methods aim to recover scenes from available data, generative reconstruction goes a step further—synthesizing realistic novel views that extend beyond the captured input. In this talk, I will present recent advances in generative reconstruction of dynamic 3D scenes, spanning a range of inputs from sparse-view images to text prompts. I’ll highlight key ideas behind feed-forward reconstruction pipelines and video generation models that handle large-scale dynamic environments. The talk will feature several representative works, including BTimer, STORM, InfiniCube, and Gen3C, each showcasing unique approaches to scalable, high-fidelity 3D generation/reconstruction in the wild.

Bio

Jiahui Huang is currently a senior research scientist at NVIDIA Toronto AI Lab led by Sanja Fidler. He received his Ph.D. in 2023 from the Graphics and Geometric Computing Group at Tsinghua University, China, advised by Shi-Min Hu. He was also a visiting researcher in the Geometric Computing group at Stanford University, led by Leonidas Guibas. His primary research interest lies in the joint field of 3D computer vision and graphics, including neural reconstruction, dynamic scene perception, and SLAM.


  1. Image from InfiniCube(https://research.nvidia.com/labs/toronto-ai/infinicube/). ↩